The Linux Foundation

LF Live Webinar: Agentic AI in the Wild: Rethinking Trust When Your AI Has the Keys

The webinar examined the emerging security challenges of agentic AI—autonomous systems that act without human oversight—and why traditional perimeter‑based defenses no longer suffice. Panelists from Nvidia, Intel and Agile Systems argued that confidential computing, which protects data while it is being processed, must become a core component of enterprise security architectures. Key insights included the inadequacy of encrypt‑at‑rest alone, the emergence of new attack surfaces such as prompt injection, and the need to treat data as executable code that requires continuous verification. Jesse highlighted that once data is decrypted for use, it is exposed in memory, while Daniel emphasized attestation, signed containers, and verifiable workloads as mechanisms to maintain trust over time. Phix underscored the risk of inference providers seeing raw data and suggested on‑prem, encrypted remote execution, or hardware‑enforced confidential AI as mitigation paths. The discussion also stressed that trust must be negotiated across the entire stack—from silicon‑level hardware roots to cloud hypervisors and model providers. Building a chain of provenance through hardware‑based attestation and policy‑driven decisions enables organizations to meet data‑location and regulatory requirements while still leveraging powerful autonomous agents. For enterprises, the implication is clear: security strategies must evolve to incorporate confidential computing primitives, continuous evidence generation, and ecosystem‑wide standards. Without this shift, the promise of agentic AI—greater productivity and real‑time decision making—remains constrained by compliance risk and potential data breaches.